Figure 1. GBM-Locfit implementation consists of two distinct steps.

First, the iteration threshold is determined per the risk criterion, by running the GBM loop within a 10-fold cross-validation loop, in order to estimate the prediction mean squared error and associated standard error for each iteration. Second, the final GBM-Locfit model is fit with 100% of the data, while limiting the number of GBM iterations to the iteration threshold. This approach utilizes all of the available data for training, gives equal consideration to each compound, and avoids over-fitting the model to the data.